China develops a social simulation system with more than 1 billion AI agents

(Dan Tri) - Light Society can simulate more than 1 billion AI agents, each with its own demographic profile, personality, memories, beliefs and goals to study how social behavior forms and spreads.

A group of Chinese researchers developed the Light Society social simulation system, capable of creating a virtual society with more than 1 billion AI agents.

Each agent not only executes pre-programmed rules but is also assigned human-like characteristics such as age, education level, income, personality, memories, beliefs, goals, location and social relationships.

The research titled "Modeling Earth-Scale Human-Like Societies with One Billion Agents" was posted on arXiv and last updated on June 28.

The authors say Light Society is designed to simulate social phenomena at a very large scale, thereby supporting the study of how beliefs, opinions and behaviors change when individuals interact with each other.

Simulate over 1 billion AI agents, each with their own memories and beliefs

The system is called Light Society, introduced in the study "Modeling Earth-Scale Human-Like Societies with One Billion Agents" (Photo: Arxiv).

Unlike traditional agent models that are mainly based on fixed rules, Light Society describes each agent with many layers of information.

Among them, relatively stable profiles include characteristics such as demographics and personality. States that can change over time include memories, beliefs, and goals. In addition, the system also tracks external factors such as location and social network.

Large language models (LLMs) are used to help agents receive information, choose actions, and update status after interactions.

However, if we use a large language model for each decision of billions of agents, the computational cost will be huge. This is the main problem that Light Society seeks to solve.

The system uses a combination of many types of models. Complex tasks can be handled using LLM, while iterative decisions are passed to smaller models that have been trained to partially simulate LLM behavior.

The research team also used caching for repeated queries, compressing the social network, processing data in batches, and gathering similar events to execute at the same time.

According to the study, this approach helps Light Society scale simulations to more than 1 billion agents.

In terms of structure, the system consists of three main components: agents, environment and event queue.

Agents interact with the environment and with each other. Each interaction is put into an event queue, which is then processed by the system to update the state of the agents or environment.

Thanks to this, Light Society can simulate interactions between two individuals, such as trust games, to the spread of opinions on a network of billions of nodes.

AI simulates trusting and reciprocating behavior

To evaluate the system, the research team performed two large groups of experiments, including simulating trust behavior and simulating opinion spread.

In the first experiment, agents participated in the "Trust Game", a form of economic game used to study trust and reciprocity between individuals.

Each player is simulated with a profile based on actual data from Wave 7 of the World Values ​​Survey. Characteristics such as age, country, and social status are used to create a natural language profile for the agent.

Simulation results show that agents belonging to groups with higher social status or education tend to deposit more money in the game, which represents a higher level of trust.

The team also found that response behaviors varied with the amount of money transferred. However, the level and specific results may vary depending on the language model used.

As the simulation scale increases, random fluctuations at the individual level decrease, while some differences between demographic groups become more pronounced.

This is the result of a simulation, not evidence that people with high levels of education or social status behave this way in real life.

Simulate how people believe and change their opinions

AI simulates how humans believe and change their opinions (Photo: News9live).

The second experiment focuses on opinion diffusion in a network of 1 billion actors.

The research team built a network with a structure similar to social networks, in which some actors have a lot of connections and are considered "influencers".

Actors were exposed to different perspectives on a controversial issue: whether AI automation will lead to widespread unemployment.

After several rounds of interaction, the research team tracks how the change agents' perspectives change.

Simulation results show that the initial opinion of an influential group can significantly impact the overall opinion trend of the network.

When the influential group leans in one direction, the views of the simulation community also tend to shift in that direction. On the contrary, when this group's views are more balanced, the system tends to be less polarized and shifts toward a neutral state.

The team also looked at the influence of individual characteristics. In the simulation, agents with higher education or income tend to persuade others more easily but at the same time are less likely to change their views.

These results suggest that network size and an individual's position in the network can influence how information spreads. However, this is still the result of a computer model, not a direct measurement of real-life human behavior.

Not a tool to predict the future

According to the research team, Light Society can be used as a "virtual laboratory" to test hypotheses in social sciences.

Researchers can change certain conditions in a simulated environment and then observe how different groups of agents react. In the long term, this approach could support policy research, information diffusion, collective behavior or other social phenomena.

However, Light Society is not a system that reliably predicts the future.

The simulation results depend on the data used to create the agent, the network structure, the model rules, and the language model used. LLM itself can also carry biases from the training data.

Therefore, the team views Light Society primarily as a tool for generating and testing hypotheses, rather than using simulation results as direct evidence of how humans would behave in reality.

The team says future developments include improving the agent's long-term learning and memory capabilities, simulating behavior over longer periods of time, and increasing the level of connectivity with real-world data.